Re: Superblock of raid5 log can't be updated when array stoped

[Shaohua Li <shli@kernel.org>]

On Sat, Oct 15, 2016 at 10:19:36AM +0800, liuzhengyuan wrote:
> Hi, Shaohua.
> 
> when we stop raid5 array with "mdadm -S" or reboot the system,  md module will 
> call raid5_quiesce and r5l_quiesce to do some clean work. Some code of r5l_quiesce
> was pasted bellow.
> 
>                 /* make sure r5l_write_super_and_discard_space exits */
>                 mddev = log->rdev->mddev;
>                 wake_up(&mddev->sb_wait);
>                 r5l_wake_reclaim(log, -1L);
>                 md_unregister_thread(&log->reclaim_thread);
>                 r5l_do_reclaim(log);
> +                md_update_sb(mddev, 1);
> 
> It will reclaim all used space of log and call r5l_write_super to reset rdev->journal_tail
> to log->next_checkpoint . However, new rdev->journal_tail would not be written to 
> journal device for persistent because journal device may not support discard operation
> or due to mddev_trylock fail (this trylock should always get failed since raid5_quiesce 
> was called with  reconfig_mutex hold, isn't it?). As a result, it will take a long time to
>  recovery the log when the arrary was restarted. Should r5l_quiesce call md_update_sb
>  directly to guarantee  superblock  update?

Yep, that's problem here. Unfortunately we can't call md_update_sb here,
because we might not hold the mddev lock. I think we should call it at do_md_stop.

Thanks,
Shaohua